Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 www.hydrol-earth-syst-sci.net/15/1713/2011/ Hydrology and doi:10.5194/hess-15-1713-2011 Earth System © Author(s) 2011. CC Attribution 3.0 License. Sciences

Hydrological differentiation and spatial distribution of high altitude wetlands in a semi-arid Andean region derived from satellite data

M. Otto1, D. Scherer1, and J. Richters2 1Technische Universitat¨ Berlin, Department of Ecology, Chair of Climatology, Rothenburgstraße 12, 12165 Berlin, Germany 2Lohmeyer Consulting Engineers GmbH & Co. KG, Karlsruhe, Germany Received: 4 November 2010 – Published in Hydrol. Earth Syst. Sci. Discuss.: 28 January 2011 Revised: 21 April 2011 – Accepted: 3 May 2011 – Published: 31 May 2011

Abstract. High Altitude Wetlands of the (HAWA) between temporal HAWA and precipitation (r2: 0.75) during belong to a unique type of wetland within the semi-arid austral summer (March to May). Annual changes in spatial high Andean region. Knowledge about HAWA has been extend of perennial HAWA indicate alterations in annual wa- derived mainly from studies at single sites within different ter supply generated from snow melt. parts of the Andes at only small time scales. On the one hand, HAWA depend on water provided by glacier streams, snow melt or precipitation. On the other hand, they are sus- 1 Introduction pected to influence hydrology through water retention and vegetation growth altering stream flow velocity. We derived High Altitude Wetlands (HAWs) are situated in many moun- HAWA land cover from satellite data at regional scale and tain regions of the world (e.g., Himalaya or Alps). HAWs analysed changes in connection with precipitation over the can be described as areas of swamp, marsh, meadow, fen last decade. Perennial and temporal HAWA subtypes can or peatland whether natural or artificial, perennial or tempo- be distinguished by seasonal changes of photosynthetically rary, with water that is stagnant or flowing, fresh, brackish, active vegetation (PAV) indicating the perennial or temporal or saline (Chatterjee et al., 2010). Although there is no scien- availability of water during the year. HAWA have been delin- tific definition of HAWs, we would use the following general 2 eated within a region of 12 800 km situated in the Northwest working definition: HAWs can be seen as any kind of tem- of . The multi-temporal classification method porarily or perennial water saturated ground above the nat- used Normalized Differenced Vegetation Index (NDVI) and ural forest border and below the snow line within any high Normalized Differenced Infrared Index (NDII) data derived mountain region on earth. from two Landsat ETM+ scenes at the end of austral winter HAWs of the arid and semi-arid region of the Andes (September 2000) and at the end of austral summer (May (HAWA) are situated in environments of relatively low an- 2001). The mapping result indicates an unexpected high nual precipitation and soil moisture deficits within a region 2 abundance of HAWA covering about 800 km of the study also known as Puna (Wilcox et al., 1986). HAWA exist at hy- region (6 %). Annual HAWA mapping was computed using drological and altitudinal limits for plant life in the cold and NDVI 16-day composites of Moderate Resolution Imaging arid high Andean grasslands of Peru, Bolivia, Chile and Ar- Spectroradiometer (MODIS). Analyses on the relation be- gentina. In the central parts of the Andes (16◦ S), they occur tween HAWA and precipitation was based on monthly pre- between 4000 m a.s.l. and 5000 m a.s.l., and in the southern cipitation data of the Tropical Rain Measurement Mission parts (27◦ S) between 2000 m a.s.l. and 3000 m a.s.l. (Scott (TRMM 3B43) and MODIS Eight Day Maximum Snow Ex- and Carbonell, 1986). The Convention on Wetlands of Inter- tent data (MOD10A2) from 2000 to 2010. We found HAWA national Importance (Ramsar Convention) compiled a list of subtype specific dependencies on precipitation conditions. A wetlands, which includes HAWA due to their importance for strong relation exists between perennial HAWA and snow fall sustaining biodiversity including endemic flora and fauna of 2 (r : 0.82) in dry austral winter months (June to August) and the Puna region. HAWA also play a critical role for local live- stock grazing of the Andean camelid species like wild Vicuna˜ (Vicugna vicugna), (Lama guanicoe) and domes- Correspondence to: M. Otto ticated forms like (Vicugna pacos) or Lama (Lama ([email protected]) glama)(Moreau et al., 2003). Knowledge about HAWA has

Published by Copernicus Publications on behalf of the European Geosciences Union. 1714 M. Otto et al.: High altitude wetlands in a semi-arid Andean region Bolivia

!( "

! Lake Anan ta Juliaca ! !( Chivay !( ! Lake Titicaca (! Puno (!

! !( (! River " Arequipa CHIVAY ! Coata Peru " !( !( " !( "

" G " " Imata " " ! 0 30 60 120 km " *# B G """ 70°W 60°W " Eston Sumbay G !

!( Brazil *# !( *# " " S S ° °

0 !( 0

1 1 G !(!(!( "" Peru !( "" !( !( !( " !( !( !( !( !( AREQUIPA Lake Salinas !( !( ! " !(!(!( Bolivia VITOR ! " !(!( Study area (4000 m a.s.l.) !( S S ° °

0 !( !( 0 Transect 2 P a c i f i c 2 !(!( Paraguay ") HAWA(gt) !( !( O z e a n !( !( Grasslands(gt) !( !( Chile Argentina *# . A Shrublands(gt) !( G barren or sparsely vegetated(gt) !( meters a.s.l. 70°W 60°W 0 5 10 20km !( HAWA sites (Source: Scott & Carbonell, 1986) C 4000 5000

Fig. 1. Location of the study region within the known geographic range of HAWA (A, B) and ground-truth data (gt) above 4000 m a.s.l. (C). been derived mainly from botanical studies within different land cover data of different HAWA subtypes and precipita- parts of the Andes (e.g., Ruthsatz, 1993). HAWA depend tion data. on water provided by glacier streams, snow melt or pre- Since it is difficult to obtain spatially and temporally con- cipitation. HAWA are also suspected to influence regional sistent information on HAWA due to their often isolated and hydrology due to their water retention capacity and vegeta- remote locations, a HAWA mapping procedure was designed tion growth altering stream flow velocity (Earle et al., 2003). for the utilisation of vegetation remote-sensing data. The po- However, little is known about the relation between regional tential for continuous monitoring of HAWA applying remote- climate and hydrology of HAWA on the background of pro- sensing data from Moderate Resolution Imaging Spectro- jected changes in temperature and precipitation within the radiometer (MODIS) was quantitatively assessed. Finally Andes (Vuille et al., 2008). HAWA land cover change was computed for each year be- Knowledge on spatial variability of HAWA is non- tween 2000 and 2010. The resulting time series of annual systematic and incomplete (Naranjo, 1995). A new HAWA HAWA land cover was used to analyse relations between pre- land cover dataset is needed for investigating relations be- cipitation, snow cover and spatial alterations of the HAWA tween precipitation and vegetation changes as a proxy for subtypes. hydrological differentiation of HAWA. Thus, the first objec- tive of the study was to determine spatial extent, distribu- Study region tion and composition of HAWA within a subregion of their known geographic range. For this reason, we first devel- The study region is located on the western slopes of the oped a HAWA mapping design incorporating HAWA sub- South Peruvian Andes between 15◦150 and 16◦350 S lati- types based on specific vegetation changes and changes in tude and 70◦510 and 71◦520 W longitude covering a total water availability as consequences of the main hydrological area of approx. 12 800 km2 above 4000 m a.s.l. within the processes. The relationship between water sources and veg- semi-arid high Andean mountains. The study region was etation dynamics of HAWA has not yet been investigated in selected because of its central position within the known detail (Squeo et al., 2006). Hence, the second objective of the HAWA geographic range (Fig. 1a) containing different types study was focused on analysing this relation based on annual of the Puna region characterised by the differences in annual

Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 www.hydrol-earth-syst-sci.net/15/1713/2011/ M. Otto et al.: High altitude wetlands in a semi-arid Andean region 1715 precipitation and duration of humid season (wet season). differences in seasonality. Satellite instruments measure so- Most precipitation occurs in austral summer between Octo- lar radiation reflected by vegetation in certain spectral bands. ber and May. The north-eastern part belongs to the moist Spectral reflectance data can be converted into vegetation in- Puna belt northwest of Lake Titicaca (Fig. 1a and b) with dices. Many of these indices correlate well with e.g., veg- annual precipitation of more than 500 mm and five to six hu- etation amount, the fraction of absorbed photosynthetically mid months. The south-western part of the study region be- active radiation (PAR) or unstressed vegetation conductance longs to the dry Puna belt where mean annual precipitation and photosynthetic capacity (Myneni et al., 1995). We refer is lower, varying between 300 mm and 500 mm during three to the term Photosynthetic Active Vegetation (PAV) as land to four humid months (Baied and Wheeler, 1993; Richter, cover containing unstressed vegetation of high photosyn- 1981; Troll, 1968). thetic capacity enabling an efficient use of PAR. Therefore, Normalized Difference Vegetation Index (NDVI) was used as a measure of PAV based on spectral change in reflectance 2 Data and methods for HAWA mapping caused by chlorophyll absorption in red wavelengths (ρRed) and leaf additive reflectance in near-infrared wavelengths 2.1 Pre-processing of remote-sensing data (ρNIR)(Jackson and Huete, 1991). Other studies applying Landsat ETM+ data has been widely used for the investi- NDVI for biomass assessment of HAWA in Bolivia demon- gation on wetlands (e.g., Ozesmi and Bauer, 2002). For strated their suitability for vegetation analysis taking into ac- mapping of HAWA in dry and in wet season, we utilized count the low atmospheric effects due to high altitude, and two cloud-free, terrain corrected Landsat ETM+ datasets at the low influence of soil background due to dense HAWA 30 m spatial resolution (L1T) of 28 September 2000 and of vegetation cover (Moreau et al., 2003). The NDVI has been 5 May 2001(provided by the US Geological Survey website). known for its strong relation to precipitation. This relation Atmospheric correction was not applied since the compari- could even be applied to downscale precipitation data of lower spatial resolution utilizing NDVI data of higher spa- son of spectral index 5th percentile (P5%) and of 95th per- tial resolution (Immerzeel et al., 2009). For this study, NDVI centile (P95%) thresholds of bare soil (B & S) land cover ground-truth data revealed insignificant differences between values were derived from near-infrared band (ρNIR) and red band (ρRed) by: Landsat ETM+ (P5%: 0.02, P95%: 0.06) and MODIS data (P5%: 0.03, P95%: 0.06). ρ − ρ NDVI = NIR Red . (1) The spectral response of wetlands has shown to be similar ρNIR + ρRed to the spectral response of irrigated crop lands, which makes it difficult to differentiate both land covers (Ozesmi and A second spectral index, i.e., the Normalized Difference Bauer, 2002). Lower altitude limit of HAWA (4000 m a.s.l.) Infrared Index (NDII), was applied to discriminate parts within the study region is equal to the maximum altitude limit dominated by water surfaces or water oversaturation within for most common Andean crops (Tapia, 2000). Hence, the HAWA. The NDII is more correlated with canopy moisture study region was delineated applying this altitude limit us- than the NDVI due to the spectral response of water and soil ing the Digital Terrain Model (DTM) derived from data of moisture in the short-wave infrared band (ρSWIR)(Hardisky the Shuttle Radar Topography Mission (SRTM Version 3). et al., 1983). The NDII has the ability of retrieving vegetation The DTM was also applied for generating a shadow mask water content at 30 m spatial resolution and was computed by based on the sun elevation and azimuth values of each Land- Davidson et al. (2006): sat ETM+ scene. ρ − ρ NDII = NIR SWIR . (2) The HAWA mapping procedure, based on the devel- ρNIR + ρSWIR oped HAWA mapping design, was analogously applied to a MODIS spectral index dataset derived from daily surface re- 2.2 Ground-truth data acquisition flectance data (MOD09 and MOD02) for the same days as for Landsat ETM+. Landsat ETM+ and MODIS have differ- HAWA ground-truth sites were documented during field vis- ent spatial resolutions as described by Chander et al. (2009) its obtaining GPS coordinates and photographs (as e.g., in (Landsat ETM+) and Barnes et al. (1998) (MODIS). In or- Fig. 3) along two transects (Fig. 1c) within the study region der to quantitatively assess the HAWA mapping results, the at the end of austral winter in September 2006. Ground-truth MODIS input data was converted to spatial resolution of the data of four HAWA sites, three Grassland (Gra) sites, two Landsat ETM+ grid applying bilinear interpolation. Shrubland (Shr) sites and three sites of Barren or Sparsely vegetated ground (B & S) were collected recording GPS way- Spectral indices points within or close to homogeneous areas during the field trips. Spatial extent of each ground-truth site was digitized Spectral indices were computed to differentiate HAWA sub- as regions of interest (ROI) through visual selection of ho- types and discriminate between HAWA and other land covers mogeneous areas close to recollected GPS coordinates using through thresholds based on their absolute differences and a false colour image of Landsat ETM+ of September 2000. www.hydrol-earth-syst-sci.net/15/1713/2011/ Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 1716 M. Otto et al.: High altitude wetlands in a semi-arid Andean region

bare ground & To minimize mapping errors during visual digitalisation of HAWA T HAWA P open water open shrubs the ground-truth data, five hundred randomly sampled grid values were computed from ground-truth ROI-mask for each HAWA T,tf HAWA P,tf, land cover class (HAWA(gt), Gra(gt), Shr(gt) and B & S(gt)). The resulting new mask was used to compute spectral index thresholds for each land cover and remote-sensing dataset. Besides the ground-truth data obtained during the field vis- pools, rivulets, rivers or lakes perennial PAV temporal PAV its, we utilized additional information from different sources (cushion-plants) (grass vegetation) to generate independent HAWA test sites (Fig. 1c) as a ba- sis for formal accuracy assessment of the HAWA mapping Fig. 2. Schematic cross-cut through an ideal HAWA (modified after result. Land cover inventories like Topographic Map sheets Ruthsatz, 2000). of the Peruvian national topographic map (TMP) (IAI, 2010) and the natural resource map (INRENA, 2002) contain infor- mation on potential HAWA sites at regional scales. Scott and (Ruthsatz, 1993). The Ramsar Convention defines HAWA Carbonell (1986) incorporated information on HAWA geo- as inland wetlands that are fed by waters from snowmelt or graphic distribution for the neotropical region at continental rain, and as non-forested peatlands, including shrubs, open scale (see Fig. 1a). The Ramsar list of wetlands of interna- bogs, swamps and fens (Wetlands International, 2010). TMP tional importance contains information on reported HAWA defines HAWA as swamps, but without further differentia- locations and properties within a global wetland inventory tion to other types of wetlands, for example, coastal wetlands (Wetlands International, 2010). These datasets where uti- (IAI, 2010). lized to identify about 600 potential HAWA test sites within HAWA vegetation is sometimes compared to Northern the study region. As a decision criterion, we only selected Hemisphere peatlands suggesting that peat development can potential HAWA sites as test sites if they were mapped in at be rapid and extensive, although other plants than Sphagnum least two of the four above mentioned sources. As a result, sp. are involved (Earle et al., 2003). Most investigations on 19 HAWA test sites were digitalised computing a HAWA test HAWA provide information on floristic composition based site ROI-mask based on the Landsat ETM+ false colour im- on single sites in Peru, Bolivia, Chile and Argentina (Ruth- age. The final HAWA test site dataset consists of a thou- satz, 1993, 2000). sand randomly stratified sampled grid values derived from Figure 2 depicts a schematic cross-cut view of an ideal the HAWA test site ROI-mask. HAWA site. Dense PAV cover within HAWAP could consist A HAWA mapping criterion is also based on an altitude of cushion-plants interrupted by shallow pools, small pud- threshold regarding altitude limits for irrigated crop lands. dles or rivulets depending on climatologic conditions that Therefore, way points were recorded during field visits rep- are spatially or temporarily variable (Coronel et al., 2004). resenting ground-truth sites along two transects of altitudinal HAWA can be situated along rivers, lakes (open water in gradients from 5000 m a.s.l. to 1500 m a.s.l. (Fig. 1c). Thus, Fig. 2) or below springs (spring water mire) and gradually the information on altitudinal thresholds for HAWA found in or abruptly change from perennially dense cover of PAV literature could be validated for the study region based on (HAWAP) to vegetation of just temporarily high fractions of ground-truth information. PAV (HAWAT). HAWA sites can be surrounded by either grasslands, open shrubs or bare grounds. 2.3 Mapping of HAWA Figure 3 depicts different HAWA sites within the study re- gion at the end of austral winter (September 2006). Veg- 2.3.1 The concept of HAWA subtypes etation and water surface cover can be different from site to site ranging from typical dense cushion-plant cover inter- Based on a literature review, an evaluation of existing land rupted by pools or rivulets (Fig. 3 top right and lower left) cover inventories, field visits of ground-truth sites and fol- to totally dried out sites containing annual grasses (Fig. 3 lowing our general definition of HAWs, we primarily defined top right). Dense cushion-plant cover of HAWA sites de- two main HAWA subtypes: Perennial HAWA (HAWAP) and pends on hydrologic conditions allowing constant water sur- temporal HAWA (HAWAT). The differentiation is based on plus and, therefore, are limited to areas, for example, valley the link between temporal or perennial PAV and temporal bottoms of constant river discharge as depicted in Fig. 3 top or perennial availability of water. Unlike temporal HAWA, left and lower right (Ruthsatz, 1993). Cushion-plant dom- perennial HAWA contain constantly high fractions of PAV inated HAWA parts are also known as Bofedales containing within a hydrological year starting 1 June and ending at over 60 plant species with short grasses and dwarf reeds from 31 May. few millimetres to few centimetres tall (Moreau et al., 2003). HAWAP have been described in literature as minerotrophic Within Bofedales, species of the Juncaceae Oxchloe andina biotopes containing parts of dense cushion-like vegeta- or Distichia muscoides and Patosia clandestina can be found tion cover interrupted by pools and superficial rivulets at their whole geographic range forming extensive perennial

Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 www.hydrol-earth-syst-sci.net/15/1713/2011/ M. Otto et al.: High altitude wetlands in a semi-arid Andean region 1717

HAWA

river HAWA T HAWA P HAWA T,nf HAWA p,nf

HAWA T,tf HAWA P,tf

Grassland/ open Shrubs small lake

river

Grassland/ open Shrubs

small lake

Fig. 4. Left side Landsat ETM+ image in false colour mode (band order 4, 3, 2 in RGB) of ground-truth site upper River Coater (left side) together with corresponding NDVI (solid line) and NDII (dashed line) transects (diagrams on the right side) of Septem- Fig. 3. Photos taken from different HAWA sites at the end of the ber 2000 (above) and May 2001 (below). Grey frames within in dry season in 2006. Top left: dense cushion-plant cover of HAWA. the diagrams indicate different HAWA subclasses as described in Top right: Dried out HAWA site only containing small fractions of the text. cushion-plant cover. Middle left: HAWA site in about 5000 m a.s.l. Middle right: HAWA site containing dense cushion-plant cover right in front of a small cascade. Lower left: same site as lower 2.3.2 Development of the HAWA mapping procedure right, but photographed from elevated position, note contrast be- tween dense cushion-plant cover and surrounding bare ground or The transformation of HAWA subtypes into HAWA sub- shrub cover (small white dots within HAWA site indicate a grazing classes utilizing remote-sensing data was the main task of alpaca-herd). the HAWA mapping procedure (Fig. 4). The HAWA map- ping procedure incorporated a classification method in a step- wise manner. A variety of classification methods for remote- vegetation covers within Andean grasslands or shrublands sensing data exist ranging from visual interpretation over un- (Ruthsatz, 2000). supervised and supervised classifiers to more sophisticated Dense cushion-plant cover of HAWAP can also influence methods involving principal component analysis, neural net- local hydrologic conditions through changing water flow works or fuzzy logic approaches. The more sophisticated paths and discharge velocity. As, an example, Fig. 3 (middle methods have the advantage of handling complex relations right) shows a HAWA site in front of a small cascade mean- between classes. Their major disadvantage is that they re- dering into the green and dense cushion-plant cover. Within quire substantial expertise and human interaction to fully uti- HAWAP there are areas of temporal water oversaturated con- lize their potentials. Decision Tree classifier (DT) have many ditions (e.g., Fig. 3 top left) forming shallow pools, rivulets of the advantages of the sophisticated methods but are much or even flooded areas (Coronel et al., 2004). These parts we easier to use (Friedl and Brodley, 1997). refer to as temporarily flooded HAWA (HAWAP,tf) within Therefore, a supervised multi-temporal mapping proce- HAWAP or HAWAT,tf within HAWAT. Within the study re- dure based on hierarchical structured binary DT was applied gion HAWA appear between 4000 m a.s.l. and 5000 m a.s.l. for this study. The structure of a DT classifier consists of a Figure 3 middle left depicts HAWA site at about 5000 m a.s.l. root node, intermediates and terminal nodes separating data based on a set of decision rules. The decision rules were compromised of a set of spectral index thresholds derived from ground-truth data in order to separate HAWA land cover from other land covers and then differentiate HAWA land cover into subclasses following the developed HAWA map- ping design. In the first step of the mapping procedure, DT www.hydrol-earth-syst-sci.net/15/1713/2011/ Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 1718 M. Otto et al.: High altitude wetlands in a semi-arid Andean region classifiers were used to differentiate between HAWA and depends stronger on snow cover could be decreased result- other land cover classes once for the austral winter data of ing in a different RHAWAP/T. 2000 and again for austral summer data in 2001. For this In order to investigate relations between changes in land purpose, a general land cover scheme similar to the land cover proportion of HAWA and precipitation, three-monthly cover scheme used by the International Geosphere-Biosphere means of precipitation and snow cover time series were Program (IGBP) was applied and consists of four main nat- correlated computing Pearson’s correlation coefficients and ural vegetation classes (Loveland et al., 2000): (1) Wet- regressions. Three-monthly mean precipitation rates were lands (HAWA), (2) Grasslands (Gra), (3) Shrublands (Shr), derived from monthly precipitation data of the Tropical (4) Barren or Sparsely vegetated (B & S). Secondly, a DT Rain Measurement Mission (TRMM-3B43) and ten years classifier was used to separate the four HAWA subclasses of snow cover data was derived from MODIS Eight Day from each others incorporating land cover status of wet sea- Maximum Snow Extend data (MOD10A2) and converted son (September 2000) and dry season (May 2001). Finally, into three-monthly average snow cover data. All MODIS thresholds derived from the relationship between NDII and datasets were downloaded via The Warehouse Inventory NDVI were applied to derive temporarily flooded HAWA Search Tool (WIST) provided by National Aeronautics and subclasses. Space Administration (NASA). Three-monthly mean precip- Formal accuracy assessment was performed based on the itation rates of TRMM-3B43 data were downloaded using ground-truth test site data of each land cover class calculat- the Giovanni online data system (http://disc2.nascom.nasa. ing overall and class-specific accuracies, Kappa Coefficients, gov/Giovanni/tovas/). errors of commission and omission, as well as user and pro- ducer accuracies. Data derived from HAWA test sites used for accuracy assessment were not included in derivation of 3 Results vegetation index thresholds incorporated as decision rules into the DT classifier. 3.1 HAWA mapping design

2.3.3 Annual HAWA mapping and applied precipitation The characteristics of HAWA subtypes had to be transformed data into HAWA subclasses following the HAWA mapping de- sign. The HAWA mapping design was the conceptual frame The second objective of the study was focused on hydrolog- for the HAWA mapping procedure which was applied for ical differentiation of HAWA and HAWA subtypes regarding both HAWA mappings utilizing first Landsat ETM+ and sec- precipitation as a main driving factor within semi-arid cli- ondly MODIS datasets. Figure 4 depicts an example of a mates. After assessment of MODIS based HAWA mapping HAWA site and its surroundings in false colour mode from for the year 2000 to 2001, a time series of annual HAWA land Landsat ETM+ of September 2000 (top left) and May 2001 cover was computed utilizing MODIS 16 day NDVI compos- (lower left) together with corresponding NDVI and NDII val- ites at spatial resolution of 250 m (MOD13Q1) from 2000 to ues along the black and white marked transect (Fig. 4 right). 2001 (00 01) to 2009 to 2010 (09 10). The MOD13Q1 prod- Dashed areas in Fig. 4 (left) indicate parts of one of the uct has the advantage of combining best NDVI estimates over ground-truth sites (upper River Coata). Also parts of the a 16 day time period reducing negative impacts on data qual- upper tributary of the River Coata can be identified as grey ity by, e.g., clouds (Huete et al., 2002). winding line south and south-west of the ground-truth site. The data pre-processing for the annual HAWA mapping Note the remarkable change in the red tone (high reflectance consists of two main steps. First, we derived NDVI thresh- values in ρNIR) which corresponds to changes of spectral in- olds based on HAWA(gt) from MOD13Q1 composites of dices (Fig. 4 right) of the transect part between the river and 2000 to 2001 which we then implemented into the HAWA the eastern ground-truth site in September 2000 compared to mapping procedure for all years. Second, annual land cover May 2001. proportions (in %) of HAWAP and HAWAT were com- Grey frames in the diagrams of Fig. 4 (right) indicate dif- puted from 2000 to 2010. To evaluate the relations found ferent HAWA subtypes. The outer frame marks the part between HAWA and precipitation, we computed the ratio of the transect crossing the HAWA site. The larger left (RHAWAP/T) between the annual proportions covered by the frame within the HAWA part indicates temporal HAWA two main HAWA subclasses as follows: (HAWAT) because both vegetation indices show higher tem- poral variation compared to the part crossing perennial HAWAP RHAWAP/T = . (3) HAWA (HAWAP) which is marked by the smaller right frame HAWAT in the diagrams of Fig. 4. Within HAWAT and HAWAP RHAWA P/T should be sensitive to changes in land cover there are areas (small grey frames in Fig. 4) identified as proportions of each HAWA subclass. In some years when, HAWAT,tf and HAWAP,tf which contained higher NDII val- for example, snow cover proportion was less than in other ues in May 2001. Since the NDII reflects high propor- years land cover proportion of the HAWA subclass which tion of water surface and plant water content HAWAT,tf

Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 www.hydrol-earth-syst-sci.net/15/1713/2011/ M. Otto et al.: High altitude wetlands in a semi-arid Andean region 1719

Table 1. Three assumptions (1st–3rd) made based on characteristics in vegetation and hydrologic conditions of HAWA subtypes and derived HAWA subclasses for HAWA mapping based on satellite data as shown in Fig. 4 and introduce into the HAWA-mapping design (Fig. 5). Bold and case-sensitive letters indicate the link between the characteristics of HAWA subtypes and abbreviations used for the HAWA subclasses.

1st 2nd 3rd Spectral HAWA HAWA HAWA subtypes Hydrology Vegetation indexes subclasses

Perennial water surplus Perennially HAWA dense vegetation cover P non flooded highest NDVI HAWA of Perennial water P,nf containing cushion-plants surplus and Perennially Perennially highest PAV highest fraction of PAV Perennial water surplus highest NDVI compared to HAWA temporarily flooded Temporarily P,tf other land highest NDII cover types in temporal water surplus Temporal HAWA the Puna annual vegetation cover T non flooded highest NDVI HAWA (>4000 m a.s.l.) of temporal water annual grass T,nf surplus and temporarily temporal water surplus vegetation Temporarily HAWA highest fraction of PAV temporarily highest T,tf flooded NDVI/NDII

and HAWAP,tf are referred to as temporarily flooded ar- Puna region above 4000 m a.s.l.(16°S) low NDVI high NDVI eas within HAWA. Non-flooded areas are referred to as 1st Grasslands HAWA HAWAP,nf and HAWAT,nf. Figure 4 indicates that HAWAP,tf contained greater NDII variance compared to HAWAP,nf be- PermanentPerennially high high NDVI NDVI TemporarilyTemporal high NDVI HAWA HAWA tween September 2000 (below zero) and May 2001. P T 2nd

Following the information found in literature and the re- non flooded temp. flooded non flooded temp. flooded sults of the analyses conducted on ground-truth data HAWA HAWA P,nf HAWA P,tf HAWA T,nf HAWA T,tf subtypes can be transformed into HAWA subclasses by a dry wet dry wet dry wet dry wet 3rd one-to-one translation of the three assumptions (1st-3rd) NDVI + ++ NDVI + + NDVI 0 ++ NDVI 0 + shown in Table 1 (columns 1 to 4) to develop the HAWA NDII + + NDII 0 ++ NDII 0 + NDII 0 ++ mapping design as illustrated in Fig. 5. The HAWA land cover class contains highest NDVI values compared to other Fig. 5. HAWA mapping design for the use of remote-sensing tech- land cover types in the Puna region during the year (1st). niques transforming HAWA subtypes into HAWA subclasses. The HAWA vegetation cover represents different temporal dynamics of PAV and, therefore, can be divided into two the ground-truth (gt) data based on Landsat ETM+. Within HAWA subclasses: HAWAP subclass contains perennially each box plot additionally relative frequency distribution of highest NDVI values and HAWAT subclass contains tem- NDVI values for the entire study region are shown indicat- porarily highest NDVI values only after raining season in ing that NDVI range of HAWA (P5%–P95%) is represented austral summer (2nd). HAWAP and HAWAT could contain in the flat-tailed part of the distribution (e.g., for Septem- flooded parts representing temporarily higher NDII values ber 2000: 0.27 NDVI to 0.6 NDVI). HAWA(gt) contained compared to HAWAP,nf and HAWAT,nf (3rd). highest NDVI values up to 0.74 and greatest NDVI range in After development of the HAWA mapping design, defin- May 2001 (0.31). Minimum NDVI value for HAWA(gt) was ing HAWA subtypes and transforming them into HAWA sub- below NDVI P5% of Gra(gt) in May 2001, indicating water classes, NDVI and NDII thresholds had to be derived based saturated conditions or false selected ground-truth grid ele- on ground-truth data. Index thresholds were implemented ments. Figure 6 also depicts that P5% of B & S(gt), Gra(gt) via the DT classifiers within the HAWA mapping procedure and Shr(gt) were close together or even slightly overlapping. (Fig. 4 in Appendix A) following the hierarchical structure of P5% of HAWA(gt) indicated greater distance to the closest the HAWA mapping design. Figure 6 depicts box plots de- land cover class (P95% of Gra(gt)) in September 2000 and rived from ground-truth (gt) grid elements of each land cover May 2001. NDVI percentiles of B & S(gt) did not vary as class of September 2000 (upper left) and May 2001 (upper much as other land cover classes and, hence, represented right). Numbers on top and below the single boxes in Fig. 6 lowest content of vegetation in September 2000 as well as indicate NDVI thresholds at P95% and P5% as derived from in May 2001. www.hydrol-earth-syst-sci.net/15/1713/2011/ Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 1720 M. Otto et al.: High altitude wetlands in a semi-arid Andean region

Sep 2000 May 2001

0.74

0.60

0.37 0.43

0.27 0.20 0.19 0.20 0.14 0.06 0.13 0.07 0.13 0.09 0.02 0.02

B&S gt Shr gt Gra gt HAWA gt B&S gt Shr gt Gra gt HAWA gt

Sep 2000 May 2001

small lake small lake

river river

Fig. 6. Box plots (upper part) of five hundred stratified randomly sampled grid values for each ground-truth dataset of September 2000 and May 2001. Box range is set to P5 % and P95 % together with median and maximum/minimum grid values for each land cover class. Lower part of Fig. 5 depicts HAWA mapped in September 2000 and May 2001 applying thresholds as depicted in the respective box plots (upper part).

As depicted in Fig. 6 the P5% of each land cover ground- in May 2001. The HAWA mapping results of September truth NDVI dataset were applied as thresholds separating 2000 and May 2001 based on Landsat ETM+ have class each land cover from the lower neighbouring land cover specific accuracies of above 90 % for all land covers result- class. Except for September 2000 (Fig. 6 upper left) where ing in a high overall accuracy of 93 % in September 2000 lower percentile of Grassland (P5%) was slightly overlap- and 94 % in May 2001. Kappa Coefficients are very high ping the upper percentile (P95%) of Shr(gt). Since the focus and, hence, accuracy statistics are very good for both HAWA was on mapping HAWA, we just applied the upper percentile mapping results. Percentage of User Accuracy is 100 % for (P95%) to separate Shr from Gra land cover. Since this is both mapping results which indicates a very high probability not the case for May 2001, the P5% threshold of Gra(gt) was that HAWA test site data represent HAWA land cover class. applied. B & S(gt) shows only small changes in P95% due One has to consider that all HAWA test sites were indepen- to low vegetation content. Therefore, P95% was applied as dent because they were not incorporated into NDVI threshold threshold for mapping of Shr in September 2000 as well as in derivation for the DT classifier. May 2001. P5% of B & S(gt) was used to separate B & S. All Note that in Fig. 6 (right) almost the entire area between values below P5% of B & S(gt) are not classified. Figure 6 the river and the ground-truth HAWA site were classified as (lower part) depicts a subset of the HAWA mapping result HAWA in May 2001. These areas were mapped as HAWAT (same subset as in Fig. 4) after applying NDVI thresholds as subclass applying a second DT classifier for the final HAWA decision rules to the DT classifier for September 2000 (Fig. 5 mapping result. With the above described first two steps of lower left) and May 2001 (Fig. 5 lower right). As a result HAWA mapping we could first, separate HAWA from other area of HAWA site mapped in September 2000 has almost land cover classes and secondly, HAWA land cover class doubled in May 2001, due to the fact that some HAWA parts could be discriminated into HAWAP and HAWAT subclasses. changed from e.g., Grassland in September 2000 to HAWA

Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 www.hydrol-earth-syst-sci.net/15/1713/2011/ M. Otto et al.: High altitude wetlands in a semi-arid Andean region 1721

In a third step, a threshold has to be derived based on 1 NDVI and NDII relationship to discriminate between flooded and non-flooded HAWA parts defined as HAWAP,nf and increasing plant water content HAWAP,tf or HAWAP,nf and HAWAP,tf, following the 3rd or soil moister increasing water surface cover assumption of the HAWA mapping design (Fig. 5). NDVI P and NDII are normalized indices enabling comparison of dif- %50 max P ferent land covers and also comparison of the indices them- %50 P %50 P %50 P selves. NDVI and NDII represent vegetation cover through %50 NDVI

P %50 high reflectance in ρNIR. NDII also represents plant water P %50 increasing PAVcover content or water saturated conditions (flooding) through dif- decreasing PAV cover P %50 P ferential sensitivity of ρNIR and ρSWIR caused by radiative %50 P P %50 characteristics of water within this range of the electromag- %50 netic spectrum (Tucker, 1979). Areas of dense vegetation 0 cover are represented by high NDVI values. The abscissa in 0.00 1.01 Fig. 7 (top) represents NDVI between 0 and 1. An increas- NDII ing NDVI value is a result of increasing PAV cover. In re- turn, areas of open1 water usually found within HAWAP,tf and HAWAT,tf decrease proportion of PAV cover causing linger- ing or decreasing NDVI values, but increasing NDII values increasing plant water content (along the ordinate in Fig.or soil7). moister Based on thisincreasing relation water NDII surface cover pixel values of HAWA land cover were evenly clustered into P 0.1 intervals. Then statistics of corresponding HAWA%50 NDVI max P %50 grid elements were computed for eachP %50 0.1 NDIIP interval%50 to P %50 determine the NDIINDVI threshold used for mapping of HAWAP,tf P P %50 and HAWA . As depicted in%50 Fig. 7 each NDII interval is

T,tf increasing PAVcover decreasing PAV cover represented by a box of 0.1 NDII units. Upper and lower box-P %50 P %50 borders mark P10% and P90%, respectively, for all mapped P P %50 HAWA NDVI grid element%50 values found within their corre- sponding 0.1 NDII0 interval. Horizontal bold lines within the small boxes represent0.00 the median NDVI. NDII interval of 1.01 NDII 0.45 to 0.55 represents highest median NDVI. At this inter- val HAWA NDII values still increase and NDVI retains or decreases due to the effect of increasing water and decreas- Fig. 7. Schematic representation of NDVI and NDII relations (top) and derivation of NDII threshold for HAWA computing NDVI ing PAV cover (as also depicted in Fig. 4). Finally the NDII ,tf statistics at 0.1 NDII intervals (bottom). threshold of 0.45 (red dotted line in Fig. 7 bottom) was ap- plied to differentiate between HAWAP,tf and HAWAP,nf as well as HAWAT,tf and HAWAT,nf. (1800 km2) where not classified (below 4000 m a.s.l., lakes urban areas or shadowed areas). 3.2 HAWA mapping results HAWA land cover is not equally distributed within the study region. Most HAWA sites are situated in the north- The total mapped area covers approximately 12 800 km2. ern parts of the study region. Figure 8 depicts three exam- Accuracy assessment of the final Landsat ETM+ based ples of HAWA sites from the northern part (Fig. 8, left), cen- HAWA mapping result indicates high overall accuracy of tral part (Fig. 8, centre) and the southern part (Fig. 8, right) 93 % with very good class specific accuracies for HAWA of the study region. Dashed parts within Fig. 8 represent land cover class (97 %). Therefore, HAWA mapping er- HAWA sites and different HAWA subtypes are represented ror is about 3 % applying Landsat ETM+. The area by different colours. Upper River Coata site (subset in Fig. 8 covered by HAWA is about 6 % (approximately 800 km2, left, also Figs. 4 and 6) contains large distinctive areas of 2 2 ±12 km mapping error) whereby 50 % of the HAWA area HAWAP (in total about 48 km ) and HAWAT (in total about 2 consists of HAWAP and about 50 % consists of HAWAT 58 km ). Hence, upper River Coata ground-truth site is one subclasses. Total cover of temporarily flooded HAWA of the largest HAWA sites (109 km2) within the study region. subclasses (HAWAP,tf and HAWAT,tf) is less then 1 % The upper River Coata basin has a total size of approximately (120 km2). Grasslands (Gra) cover 35 % (4500 km2), Shrub- 460 km2 (information derived from DTM) and is situated lands (Shr) about 39 % (5000 km2) and Barren or Sparsely within the moist Puna belt. About 23 % (109 km2) is covered vegetated soils (B & S) cover 5 % (800 km2). About 15 % by HAWA. This is about three times higher compared to total www.hydrol-earth-syst-sci.net/15/1713/2011/ Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 1722 M. Otto et al.: High altitude wetlands in a semi-arid Andean region

Table 2. Comparison of HAWA land cover and HAWA subtype cover mapped using Landsat ETM+ and MODIS data in a 2 × 2 contingency table. Bold numbers refer to total cover of HAWA and HAWA subclasses for each HAWA mapping result.

Landsat ETM+ TRUE FALSE TRUE FALSE TRUE FALSE HAWA HAWAP HAWAT TRUE 3.62 4.60 8.22 1.15 1.26 2.41 1.71 4.10 5.81 MODIS FALSE 2.62 89.16 1.73 95.86 1.65 92.54 6.24 2.88 3.36 HAWA,tf HAWAP,tf HAWAT,tf TRUE 0.003 0.012 0.015 0.003 0.011 0.014 0.000 0.001 0.001 FALSE 0.141 99.844 0.132 99.854 0.009 99.990 0.144 0.135 0.009

area area area Up. River Bofedales y Lag. de SE of Imata Coata (km²) (km²) Salinas (km²)

HAWA P 48 HAWA P 2 HAWA P 10 Lake Salinas HAWA P,tf 2 HAWA P,tf 0.3 HAWA P,tf 0.3

HAWA T 58 HAWA T 2 HAWA T 1

HAWA T,tf <1 HAWA T,tf <0.1 HAWA T,tf <0.1

HAWA 109 HAWA 4.7 HAWA 11 (basin size) (460) (basin size) (150) (basin size) (800)

annual PRCP (00_01): 720 mm annual PRCP (00_01): 490 mm annual PRCP (00_01): 170 mm

4000 5000

Fig. 8. Subsets of Landsat ETM+ based HAWA mapping result (only HAWA). Left: Upper River Coater, centre: South East of the village Imata, right: HAWA area of Ramsar Site Bofedales y Laguna de Salinas (Ramsar Site NO.: 1317). Numbers in attached tables indicate total cover of HAWA subtypes within corresponding basins (number top left of each table) and annual precipitation rates (annual PRCP).

HAWA cover (6 %) of the entire study region. HAWAP,tf NDVI measurements might have great potential for HAWA and HAWAT,tf cover smallest area. HAWAP,tf at upper River mapping and monitoring. Therefore, we first applied the de- 2 Coata site cover 4 % (about 2 km ) of HAWAP area which veloped HAWA mapping procedure also to MODIS NDVI is higher compared to total HAWAP,tf cover of the entire data derived from surface reflectance data sets (MOD09 and study region (less then 1 %). This is also true to HAWAT,tf MOD02) of the same time periods. MODIS and Land- covering 2 % of HAWAT of River Coata site. The HAWA sat ETM+ have different spatial and spectral characteristics. site southeast of Imata (Fig. 8, middle also in Fig. 1c) is of For this reason potential for continuous monitoring of HAWA 2 5 km spatial size but contains high percentage of HAWAP,tf applying MODIS data was quantitatively assessed in a sec- (15 %). Ramsar site Bofedales y Laguna de Salina (Fig. 8 ond step. As a result a 2 × 2 contingency table was com- right, close to Lake Salinas in Fig. 1c) is situated in the dry puted for HAWA land cover and HAWA subclasses (Table 2) 2 Puna belt and consists of about 10 km of HAWAP and about for the entire study region. The numbers in Table 2 re- 2 1 km of HAWAT. fer to area coverage in percent of total area mapped in the In order to monitor spatial changes of HAWA, data of study region. Bold numbers in Table 2 refer to total cover of better time consistency is needed. NDVI data of MODIS HAWA land cover class and HAWA subclass, respectively, (MOD13Q1) is based on daily observations and contain more for each HAWA mapping result. For HAWA land cover class than 10 years of ready to use global NDVI measurements free 3.62 % of the study region was mapped (TRUE in Table 2) available via the Internet (Huete et al., 2002). MODIS based in both HAWA mapping results. On the contrary 89.19 % of

Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 www.hydrol-earth-syst-sci.net/15/1713/2011/ M. Otto et al.: High altitude wetlands in a semi-arid Andean region 1723

20 prcp JJA prcp SON prcp DJF prcp MAM snow cover JJA 18 450 HAWA P HAWA T ± 16 400

14 350

12 300

10 250

8 200 prcp (mm) & snow& cover (%) HAWA HAWA landcover 6 150

4 100

2 50

0 0 00_01 01_02 02_03 03_04 04_05 05_06 06_0707_08 08_09 09_10 Time (years) 1 2 Fig. 9. Annual time series of MODIS based HAWA and HAWA 3 P T 4 land cover, three-monthly precipitation (mm) derived from monthly 5 rain rate of TRMM (3B43) from 1 June 2000 to 31 May 2001 6 (00 01) till 1 June 2009 to 31 May 2010 (09 10) and average three- 7 monthly (JJA) snow cover (JJA) based on MODIS Eight Day Max- 8

n seasons asn HAWAp 9 imum Snow Extend data (MOD10A2). 0500 1.000 2.000 10 Meters

Fig. 10. Annual changes in spatial extend of HAWA . Numbers the study region was not mapped as HAWA in either Land- P in coloured rectangular boxes indicate number of years where areas sat ETM+ based nor MODIS based HAWA mapping (FALSE were mapped as HAWAP. in Table 2). If we consider Landsat ETM+ based HAWA mapping as the ground-truth dataset then 6.24 % of the study region is covered by HAWA land cover class whereby 2.62 % The question is if HAWA land cover area defined in 00 01 of the HAWA site was not mapped using MODIS. Instead P was always of the same spatial extend. Figure 9 depicts land 4.6 % of the study region was mapped as HAWA only in the cover mapped as HAWA subdivided in HAWA and HAWA MODIS based HAWA mapping. Therefore, we can conclude P T indicating that spatial coverage of HAWA varies between that first, about two thirds (3.62 % TRUE in both) of Land- P 2 % and 4 % for most of the years. But some years indi- sat ETM+ based HAWA mapping (6.24 % total HAWA, bold cate higher HAWA land cover, for example, 01 02 where number in Table 2) are also mapped using MODIS. Secondly, P HAWA land cover was more than four times higher com- that MODIS overestimates HAWA site resulting in a total P pared to HAWA land cover. Figure 9 also depicts an- HAWA land cover area of 8.22 % (bold number in Table 2) T nual precipitation subdivided into three-monthly precipita- which is about 2 % higher then total HAWA land cover area tion rates (JJA, SON, DJF and MAM) and also percentage of mapped using Landsat ETM+. As shown in Table 2 overes- snow cover in austral winter months (snow cover ) for each timation primarily occurs in HAWA subclass because per- JJA T year. Both annual precipitation as well as three-monthly pre- centage of HAWA area only mapped in MODIS (4.1 %) is T cipitation and snow cover were highly variable. Annual pre- almost equal to the overestimated HAWA area (4.6 %). cipitation amount ranged from more than 400 mm in 00 01 to Quantitative assessment of HAWA subclasses indicate P,tf less than 200 mm 06 07. In general temporal distribution of that only very small fractions were mapped using MODIS. precipitation was relatively constant for the whole time pe- This is due to the general small percentage of land cover con- riod with most of the precipitation occurring during austral taining HAWA subclasses already in the Landsat ETM+ P,tf summer (60 % during DJF) and low precipitation amounts HAWA mapping result. Therefore, respective numbers in during austral winter (4 % in JJA). Snow cover in austral Table 2 are below significance and were not considered for JJA winter (six out of ten) was about 2 % or less but was signifi- further time series analysis applying MODIS-NDVI data cantly higher in 01 02 (6 %), 02 03 (16 %) and 09 10 (7 %). (MOD13Q1). Compared to Fig. 10 annual changes in spatial extend of 3.3 Hydrological differentiation of HAWA HAWAP at upper River Coata test site indicate spatial fre- quency of mapped HAWAP. The numbers in coloured rect- Taking in mind that MODIS based HAWA mapping is still angular boxes of Fig. 10 indicate the numbers of years were feasible even at coarser spatial resolution a land cover map an area was mapped as HAWAP. For the upper River Coata based on MOD13Q1 data was produced applying NDVI test site we could, therefore, specify areas constantly mapped thresholds derived from the year 2000 to 2001 (00 01). as HAWAP (maximum of ten years in Fig. 10 – dark green) Hence, the mapping criterion was left unchanged for each and areas which were not constantly mapped as HAWAP year (00 01 to 09 10) to analyse spatial changes of areas that (minimum of one year in Fig. 10 – yellow) between 2000 reached at least the same level of PAV as in 2000 to 2001. and 2010. About 2 % of the whole study region contains www.hydrol-earth-syst-sci.net/15/1713/2011/ Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 1724 M. Otto et al.: High altitude wetlands in a semi-arid Andean region

6 constantly mapped HAWAP land cover. Hence, the above 6 proposed question could be answered as follows: Not all A = 0.03e 0.05prcp HAWAP areas mapped in 00 01 were always mapped in the rA2 = = 0.75 0.03e 0.05prcp r2 = 0.75 other years. HAWAT areas were highly variable ranging from almost 5 % in 01 02 to almost 0 % in 07 08 (0.06 %). 4 4 Based on the answer of the above-stated question, we wanted to know if precipitation was a key factor for annual land cover landcover (%) T changes of HAWAP and HAWAT land cover. Figure 10 de- landcover (%) T

picts that less constantly mapped HAWA areas were at a fur- HAWA P 2 ther distance from the more constant ones indicating a spatial HAWA 2 gradient (from dark green to yellow in Fig. 10). Compared to Fig. 9 this spatial gradient might be a consequence of ab- solute and temporal changes in precipitation. Therefore, we analysed the relation between precipitation and HAWA land 0 0 0 20 40 60 80 100 120 0 20 40 60 80 100 120 cover for the entire study region. Figure 11 top depicts the prcp MAM (mm) prcp MAM (mm) dependency of HAWAT land cover to prcpMAM as well as

HAWAP land cover and snow coverJJA (Fig. 11 bottom) indi- 10 cating high correlation coefficients (r2: 0.75 and r2: 0.82, re- 10 spectively). The highest correlation coefficient was found ap- A = 0.33snow cover JJA + 2.43 8 rA2 = = 0.82 0.33snow cover JJA + 2.43 plying exponential regression for HAWAT and linear regres- 8 r2 = 0.82 sion for HAWAP land cover. Correlation coefficients of total annual precipitation (not shown in Fig. 11) and HAWA sub- 2 6 classes were lower (e.g., r : 0.6 for HAWAP) and correlation 6 coefficient based on other combinations between temporal land cover cover (%) land land cover cover (%) land precipitation means and HAWA or HAWA land cover were P T P P 4 below r2: 0.6 (e.g., for prcp and HAWA – r2: 0.0). 4 MAM P HAWA HAWA

2 2 4 Discussion

Existing HAWA mappings were of insufficient quality for in- 00 0 2 4 6 8 10 12 14 16 18 vestigations on HAWA. We used TMP combined with other 0 2 4 6 8 10 12 14 16 18 snowsnow covercover JJA (%)(%) land cover data for deriving HAWA test sites. TMP has JJA been the most detailed information source regarding distri- Fig. 11. Relation between prcpMAM and HAWAT land cover (top) bution of HAWA within the study region (IAI, 2010). About and snowcoverJJA and HAWAP land cover (bottom) of 2000 to 580 sites were classified as Swamp (Pantano) in TMP cover- 2010. ing 590 km2. In general, all ground truth sites and HAWA test sites were represented also in TMP, but were overes- timated up to 50 % and only 20 % of all mapped HAWA 2002). However, potential for continuous monitoring is lim- sites are also mapped in TMP. The number of HAWA sites ited due to restriction in data availability, we only found one in TMP is underestimated, but the area at a single HAWA set of adequate Landsat ETM+ data for the study region. site is often overestimated. The applied methodology of vi- Hence, a combination of Landsat ETM+ and MODIS NDVI sual interpretation (MULTIPLEX, A-8) based on aerial pho- data is recommended to enable time series analyses of spatial tos (black & white) of 1955 conducted in 1964 (IAI, 2010) dynamics of HAWA subtypes at different spatial and tempo- could be the underlying reason. There was no further infor- ral scales taking into account that MODIS NDVI data un- mation available on data acquisition time and the conducted derestimates HAWAP and tends to overestimate HAWAT ar- field work. The updated version of TMP of 1990 showed eas. Since this is probably due to lower spatial resolution of inconsistencies to the TMP of 1955. We found large sites the MODIS sensors, MODIS based HAWA mapping results mapped as swamps appearing only in the updated version of are systematically biased. Once this bias is quantitatively as- TMP. Compared to the information found in literature, a span sessed as shown in this study monitoring and investigation of 35 years for the development of a new HAWA site seems on HAWAT and HAWAP at spatial resolution about seventy unrealistic (e.g. Squeo et al., 2006). times lower then Landsat ETM+ was feasible. As also shown in other studies for different types of wet- As described in the HAWA mapping design HAWAP,tf and lands, Landsat ETM+ data revealed its high potentials for HAWAT,tf represent flooded areas within HAWA as gener- wetland mapping also in this study (Ozesmi and Bauer, ally expected for wetlands. HAWAP,tf and HAWAT,tf cover

Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 www.hydrol-earth-syst-sci.net/15/1713/2011/ M. Otto et al.: High altitude wetlands in a semi-arid Andean region 1725 the smallest area of the study region. Although HAWAP,tf 60 could cover up to 15 % of a single HAWA site (Fig. 8 07_08 middle). The latter could also be related to precipitation 50 because sites of relatively high precipitation contain more HAWAP,tf cover (Fig. 8 left – 720 mm and centre 490 mm) 40 than sites of lower precipitation (Fig. 8 right – 170 mm).

Flooded areas of HAWA are suspected to contain pools P/T 02_03 with mineral bottoms probably representing residual shal- 30 low open-water areas that have not yet been overgrown by RHAWA cushion-plants due to high stream velocities. Lower precip- 20 r2 = 0.62 itation rates could generate lower stream velocities allow- r2 = 0.90 03_04 ing cushion-plants to overgrow this aquatic environments and 10 04_05 cause hydroseres promoting rapid peat accumulation as sug- 09_10 05_06 01_02 00_01 gested for a single HAWA site found in northern Chile (Earle 08_09 0 06_07 et al., 2003). The proposed methodology could be used to 0 10 20 30 40 50 60 70 80 90 100 110 monitor this processes, but should be further evaluated. Un- prcp MAM (mm) fortunately, MODIS time series data did not provide reason- able results for the multi annual mapping of HAWAP,tf and Fig. 12. Relation between prcpMAM and RHAWAP/T. Grey fitted 2 HAWAT,tf cover due to its coarser spatial resolution espe- line is based on RHAWAP/T and prcpMAM for all years (r : 0.62) cially in ρSWIR (Barnes et al., 1998). and black fitted line was computed without 02 03 (highest snow 2 HAWAP depends on run off generated from snowfall as cover) (r :0.90). also suspected in other studies (e.g., Ruthsatz et al., 2003; Squeo et al., 2006). In this study dependencies of HAWAP on snow fall was quantified at regional scale for the first time. to impact of vegetation growth on surface water flow paths However, Squeo et al. (2006) stated that the morphological (Fig. 3 – middle right), range land use (Fig. 3 – lower left) or character could be important at a single HAWA site. Al- changes in peat accumulation as also discussed by Earle et though this was not within the focus of the study, spatial gra- al. (2003). dients as detected in Fig. 10 serve as an indicator to monitor plant water availability for PAV and should be investigated regarding the morphological setting within HAWA sites. 5 Conclusions The existence of constantly mapped HAWAP might in- dicate dependencies to different water sources than annual The DT classifier was used for mapping of spatial distri- snow cover or precipitation. The constantly mapped HAWAP bution of HAWA within a sub-region of the central Andes. could dependent on water sources, for example, water orig- The mapping results served as a basis for hydrological dif- inated from glaciers or even fossil aquifers. At the ecologi- ferentiation of HAWA. The study showed that DT is an ad- cal level regarding plant composition of HAWA we can con- equate method for the implementation of HAWA subtypes clude that on the one hand, the more constant HAWAP areas into a HAWA mapping procedure for the use of remote- are best suited providing optimal hydrologic conditions for sensing data. The advantage of the proposed method lies typical cushion-plant communities because they need con- in the simplicity of incorporating knowledge about HAWA stant (year-to-year) provision of water. On the other hand, based on a newly developed HAWA mapping design. There- HAWAT and less constant HAWAP should contain vegetation fore, differentiation of HAWA into HAWA subtypes as found adopted to changes in plant water availability due to annual in literature and observed during field visits was possible differences in precipitation. through transformation of HAWA subtypes into HAWA sub- HAWAT cover were controlled by precipitation occurring classes for the use of two different remote-sensing datasets. between March and May during austral summer. Changes in The developed HAWA mapping procedure is applicable to HAWAP cover were controlled by snow events during June remote-sensing data of different spatial resolution although and August in austral winter. As depicted in Fig. 12 rela- over and underestimation effects have to be considered. A tion between RHAWAP/T and prcpMAM seems to be sensitive possible strategy for investigation on HAWA within other to changes in precipitation weakened by years of high snow semi-arid high Andean regions could be to first perform a su- cover (02 03). Hence, RHAWAP/T might serve to monitor pervised multi-temporal land cover classification as proposed vegetation responds especially during years of unusual pre- in this study based on Landsat ETM+ data and ground truth cipitation events within the high Andean region. HAWA sub- data derived from existing wetland inventories (e.g., topo- types are probably not only dependent on precipitation and graphic maps combined with field data). In a second step annual changes could also be related to e.g., changing per- MODIS data could be implemented for time series analy- colation rates due to erosion, auto regulation processes due ses together with data of TRMM and MODIS snow cover www.hydrol-earth-syst-sci.net/15/1713/2011/ Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 1726 M. Otto et al.: High altitude wetlands in a semi-arid Andean region

Appendix A

Step 1

DT: NDVI September 2000 1 ( above 4000 m a.s.l.)

NDVI ge tHAWA(gt) HAWA

NDVI ge tGra(gt) Gra Step 3 NDVI NDII May 2001 NDVI ge tShr(gt) Shr Step 2 NDII ge tHAWA HAWA P,tf

NDVI ge tB&S(gt) B&S HAWA P HAWA 0 P,nf HAWA T DT: NDVI May 2001 1 (above 4000 m a.s.l.) Gra NDII ge tHAWA HAWA T,tf

NDVI ge tHAWA(gt) HAWA Shr HAWA T,nf

NDVI ge tGra(gt) Gra B&S NDVI Legend NDVI ge tShr(gt) Shr decision rule: NDVI ge tHAWA(gt) greater or equal (ge) threshold (t)

NDVI ge tB&S(gt) B&S true false 0 land cover class (if light green coloured: Gra HAWA same pixel but different land cover classes

in 2000 and 2001 representing HAWA T) Fig. A1. HAWA mapping procedure: translation of the HAWA mapping design into a decision tree (DT) classifier for the use of remote- sensing data. data to investigate spatial and temporal changes of HAWA the transitional period of austral summer to austral winter on precipitation within their whole distribution range. Since time (MAM). It must be further evaluated if these finding all remote-sensing datasets utilized for this study are free of can be concluded in other regions where HAWA occur and charge and accessible via Internet, costs for data acquisition how this processes are connected to regional climates. And are very low. finally, spatial gradients of perennial HAWA enable moni- The NDII represents vegetation of high plant water content toring of annual plant water availability for PAV at a single or surface water content reflecting differences in hydrologic HAWA site over the last decade. These findings need to be conditions within HAWA. Investigation on HAWAP,tf and further investigated at selected HAWA sites to evaluate re- HAWAT,tf applying remote-sensing data is limited to data of sults from hydrologic models with in situ measurements and higher spatial resolution (less than or equal to 30 m). Hence, remote sensing. An important question is if the observed the use of Landsat ETM+ is recommended for further inves- spatial changes have any consequences for the ecosystem tigations on HAWAft. Open water or water saturated areas functions usually related with wetlands, for example, water as defined for HAWAft are essential aquatic features of wet- storage capacity, biomass production or provision of critical lands in general and in particular within semi-arid climates habitat features for flora and fauna of the semi-arid high An- providing habitats for endemic flora and fauna of the Andes. des. Combining the potential results with regional climate The combination of the two vegetation indices provides valu- models could serve as bases to evaluate effects of climate able information on the spatial relation between the aquatic change on the regional hydrologic circle of the high Andes part (e.g., HAWAP,tf) and non-aquatic parts (e.g., HAWAP,nf) using HAWA as proxies detecting hydrological and, hence, of HAWA. ecological consequences of projected decrease in precipita- The results of the study indicate that first of all HAWA tion, increase of temperatures and increased tropical glacier are a widespread type of wetland within the semi-arid high retreat. The results of the study provide a basis for further in- Andes. The applied methodology and data has great poten- vestigation on the complexity of HAWA with regard to their tial for initiating a HAWA-inventory for these often remote hydrological functioning towards a scientifically sound long regions. Secondly, HAWA are ecosystems containing parts term management within a region of the high Andes proba- that are connected to hydrological processes through run bly facing a future scenario of decreasing precipitation and off generated from snowmelt (HAWAP). While other parts increasing water demand. (HAWAT) are more dependent on direct precipitation during

Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011 www.hydrol-earth-syst-sci.net/15/1713/2011/ M. Otto et al.: High altitude wetlands in a semi-arid Andean region 1727

Acknowledgements. We acknowledge Philipp Buss, former head INRENA: Evaluacion ambiental de la cuenca del rio Chili, Instituto of the project COPASA-GTZ in Arequipa/Peru, for providing Nacional de Recursos Naturales (INRENA), Rep., Lima, Peru, valuable support in transportation and equipment during field 2002. studies. Jackson, R. D. and Huete, A. R.: Interpreting vegetation indexes, Prev. Vet. Med., 11(3–4), 185–200, 1991. Edited by: S. Thompson Loveland, T., Reed, B., Brown, J., Ohlen, D., Zhu, Z., Yang, L., and Merchant, J.: Development of a global land cover characteristics database and IGBP DISCover from 1 km AVHRR data, Int. J. References Remote Sens., 21(6–7), 1303–1330, 2000. Moreau, S., Bosseno, R., Gu, X. F., and Baret, F.: Assessing the Baied, C. A. and Wheeler, J. C.: Evolution of high Andean puna biomass dynamics of Andean bofedal and totora high-protein ecosystems: environment, climate, and culture change over the wetland grasses from NOAA/AVHRR, Remote Sens. Environ., last 12 000 years in the Central Andes, Mt. Res. Dev., 13(2), 145– 85(4), 516–529, 2003. 156, 1993. Myneni, R. B., Hall, F. G., Sellers, P. J., and Marshak, A. L.: In- Barnes, W. L., Pagano, T. S., and Salomonson, V. V.: Prelaunch terpretation of spectral vegetation indexes, IEEE T. Geosci. Re- characteristics of the Moderate Resolution Imaging Spectrora- mote, 33(2), 481–486, 1995. diometer (MODIS) on EOS-AM1, IEEE T. Geosci. Remote, Naranjo, L. G.: An evaluation of the first inventory of South Amer- 36(4), 1088–1100, 1998. ican wetlands, Vegetatio, 118, 125–129, 1995. Chander, G., Markham, B., and Helder, D.: Summary of current ra- Ozesmi, S. and Bauer, M.: Satellite remote sensing of wetlands, diometric calibration coefficients for Landsat MSS, TM, ETM+, Wetl. Ecol. Manag., 10(5), 381–402, 2002. and EO-1 ALI sensors, Remote Sens. Environ., 113, 893–903, Richter, M.: Klimagegensatze¨ in Sudperu¨ und ihre Auswirkungen 2009. auf die Vegetation, Erdkunde, 35, 12–30, 1981. Chatterjee, A., Blom, E., Gujja, B., Jacimovic, R., Beevers, L., Ruthsatz, B.: Flora and ecological conditions of high Andean peat- O’Keeffe, J., Beland, M., and Biggs, T.: WWF Initiatives to lands of Chile between 186◦000 (Arica) and 40◦300 (Osorno) study the impact of climate change on Himalayan high-altitude south latitude, Phytocoenologia, 23, 157–199, 1993. wetlands (HAWs), Mt. Res. Dev., 30(1), 42–52, 2010. Ruthsatz, B.: Die Hartpolstermoore der Hochanden und ihre Arten- Coronel, J. S., Declerck, S., Maldonado, M., Ollevier, F., and Bren- vielfalt, Ber. D. Reinh.-Tuxen-Ges.,¨ 12, 185–234, 2000. donck, L.: Temporary shallow pools in high-Andes “bofedal” Scott, D. A. and Carbonell, M.: A Directory of Neotropical Wet- peatlands: a limnological characterization at different spatial lands, IUCN Cambridge and IWRB Slimbridge, 1986. scales, Arch. Sci., 57(2–3), 85–96, 2004. Squeo, F. A., Warner, B. G., Aravena, R., and Espinoza, D.: Davidson, A., Wang, S., and Wilmshurst, J.: Remote sensing of Bofedales: high altitude peatlands of the Central Andes, Rev. grassland-shrubland vegetation water content in the shortwave Chil. Hist. Nat., 79(2), 245–255, 2006. domain, Int. J. Appl. Earth Obs., 8(4), 225–236, 2006. Tapia, M.: Mountain agrobiodiversity in Peru: seed fairs, seed Earle, L. R., Warner, B. G., and Aravena, R.: Rapid development banks, and mountain-to-mountain exchange, Mt. Res. Dev., of an unusual peat-accumulating ecosystem in the Chilean Alti- 20(3), 220–225, 2000. plano, Quaternary Res., 59(1), 2–11, 2003. Troll, C.: The cordilleras of the tropical Americas, Aspects of cli- Friedl, M. A. and Brodley, C. E.: Decision tree classification of matic, phytogeographical and agrarian ecology, in: Colloquium land cover from remotely sensed data, Remote Sens. Environ., Geographicum, Band 9, edited by: Troll, C., Geoecology of the 61, 399–409, 1997. Mountainous Regions of the Tropical Americas, Proceedings of Hardisky, M., Klemas, V., and Smart, R.: The influence of soil salin- the UNESCO Mexico Symposium 1966, 15–56, 1968. ity, growth form, and leaf moisture on the spectral radiance of Tucker, C.: Red and photographic infrared linear combination Spartina alterniflora canopies, Photogramm. Eng. Rem. S., 49(1), for monitoring vegetation, Remote Sens. Environ., 8, 127–150, 77–83, 1983. 1979. Huete, A., Didan, K., Miura, T., Rodriguez, E., Gao, X., and Fer- Vuille, M., Francou, B., Wagnon, P., Juen, I., Kaser, G., Mark, B. G., reira, L.: Overview of the radiometric and biophysical perfor- and Bradley, R. S.: Climate change and tropical Andean glaciers: mance of the MODIS vegetation indices, Remote Sens. Environ., past, present and future, Earth-Sci. Rev., 89(3–4), 79–96, 2008. 83, 195–213, 2002. Wetlands International: Ramsar Sites Information Service, http:// IAI: Carta Nacional del Peru, scale: 1:100 000, map sheet 31s-t– ramsar.wetlands.org, last access: 11 October 2010, Wageningen, 33s-t. IGNP, Lima NIMA (&IGNP), Washington and Lima, topo- Netherlands, 2010. graphic map collection of Ibero-American Institute (IAI), Berlin, Wilcox, B., Bryant, F., Wester, D., and Allen, B.: Grassland com- 2010. munities and soils on a high elevation grassland of Central Peru, Immerzeel, W., Rutten, M., and Droogers, P.: Spatial downscaling Phytologia, 61, 231–250, 1986. of TRMM precipitation using vegetative response on the Iberian Peninsula, Remote Sens. Environ., 113, 362–370, 2009.

www.hydrol-earth-syst-sci.net/15/1713/2011/ Hydrol. Earth Syst. Sci., 15, 1713–1727, 2011